The gravitational search algorithm (GSA) is a recent addition to the family of global optimization algorithms based on phenomena found in nature, specifically the gravitational attractive force between two bodies of mass. However, like almost all global search algorithms of this type, GSA has no direct method of handling a constrained optimization problem. There has been much attention to constraint handling using other agent based systems, though the mechanics of GSA make the application of many of these difficult. This paper has therefore analysed constraint handling methods for use with GSA and compared the performance of simple to implement methods (penalties and feasible directions) with a novel separation-sub-swarm (3S) approach, and found that feasible direction methods ideally need at least one initially feasible particle, and that the novel 3S approach is highly effective for solving constrained optimization problems using GSA outperforming the other approaches tested.